Evaluation of uncertain signals’ impact on deep reinforcement learning-based real-time control strategy of urban drainage systems

Abstract

Real-time control (RTC) is a recognized technology to enhance the efficiency of urban drainage systems (UDS). Deep reinforcement learning (DRL) has recently provided a new solution for RTC. However, the practice of DRL-based RTC has been impeded by different sources of uncertainties. The present study aimed to evaluate the impact caused by the uncertainties on DRL-based RTC to promote its application. The impact of uncertainties in the measurement of water level signals was evaluated through large-scale simulation experiments and quantified using measures of statistical dispersion of control performance distribution and relative change of control performance compared to the baseline scenario with no uncertainty. Results show that the statistical dispersion of DRL-based RTC was reduced by 15.48%–81.93% concerning random and systematic uncertainties compared to the conventional rule-based control (RBC) strategy. The findings indicated that DRL-based RTC is robust and could be reliably applied to safety-critical real-world UDS.